chunk_id
stringlengths
3
7
chunk
stringlengths
1
823
source_url
stringclasses
416 values
title
stringclasses
416 values
chunk_idx
int64
0
294
chunk_start_char
int64
0
139k
chunk_end_char
int64
303
139k
32_27
We use the same methodology to classify tweets in our dataset as blackmarket or genuine. Spam Detection 2: For baseline 2, we consider the approach proposed by Rajdev et. al. BIBREF11 . They proposed flat and hierarchical classifications approaches with few of the standard set of features which can classify spam, fak...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
27
14,386
14,932
32_28
This feature vector is then fed to state-of-the-art machine learning classifiers - Random Forest (RF), Multi-layer Perceptron (MLP), and Support Vector Machine (SVM). Evaluation Setup We consider the problem as a binary classification problem, where the tweets are classified into two classes - blackmarket and genuin...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
28
14,932
15,559
32_29
The average results are reported after 5-fold cross-validation. Experimental Results As shown in Table TABREF29 , we observe that the multitask learning based model which uses the Tweet2Vec encoding and the content features as inputs to two separate tasks outperforms all the baselines, achieving an F1-score of 0.89 ...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
29
15,559
16,135
32_30
The percentage of each class in the false negatives is as follows: Promotional - 23.29%, Politics - 10.96%, Entertainment - 21.92%, News - 9.59%, Spam - 5.48%, and Others - 28.77%. We observe that the tweets belonging to the category Others are difficult to classify since they are similar to genuine tweets in terms of...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
30
16,135
16,617
32_31
Conclusion In this paper, we presented a novel multitask learning approach to solve the problem of identification of tweets that are submitted to blackmarket services, without the use of any temporal features.
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
31
16,617
16,829
32_32
To sum up, our contributions are three-fold: (i) Characterization: We proposed 12 tweet content based features that are useful in the task of identifying blackmarket tweets, (ii) Classification: We developed a novel Multitask Learning based model to classify tweets as blackmarket tweets or genuine tweets, (iii) Datase...
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
32
16,829
17,283
32_33
Acknowledgements The work was partially funded by DST (ECR/2017/00l691, DST/INT/UK/P158/2017), Ramanujan Fellowship, and the Infosys Centre of AI, IIIT-Delhi, India. Fig. 1. Architecture of our proposed multitask learning model for the detection of blackmarket tweets. TABLE II PERFORMANCE OF THE COMPETING METHODS.
https://arxiv.org/abs/1907.04072
Multitask Learning for Blackmarket Tweet Detection
33
17,283
17,603
33_0
Cross-Lingual Natural Language Generation via Pre-Training In this work we focus on transferring supervision signals of natural language generation (NLG) tasks between multiple languages. We propose to pretrain the encoder and the decoder of a sequence-to-sequence model under both monolingual and cross-lingual setting...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
0
0
599
33_1
Then the sequence-to-sequence model trained in a single language can be directly evaluated beyond that language (i.e., accepting multi-lingual input and producing multi-lingual output). Experimental results on question generation and abstractive summarization show that our model outperforms the machine-translation-bas...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
1
599
1,186
33_2
Introduction Learning natural language generation (NLG) models heavily relies on annotated training data. However, most available datasets are collected in a single language (typically English), which restricts deploying the applications to other languages. In this work, we aim at transferring the supervision of a m...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
2
1,186
1,613
33_3
Various methods have been proposed over the years to learn universal cross-lingual word embeddings BIBREF0, BIBREF1, BIBREF2 or sentence encoders BIBREF3, BIBREF4, BIBREF5, which tries to encode multilingual texts into a single shared vector space. Despite achieving promising results on cross-lingual classification p...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
3
1,613
2,227
33_4
So both encoder and decoder should be pre-trained together. Second, the many-to-many nature of cross-lingual NLG increases language pairs with the square of the number of languages. Third, the prediction space of cross-lingual NLG is much larger than classification tasks, which makes the knowledge transfer of decoders...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
4
2,227
2,741
33_5
For example, the input written in other languages is first translated to English, and fed into the NLG model that is trained by English data. Then the generated English text is translated back to the target language. Another strand of work employs MT to generate pseudo training data for other language pairs that are l...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
5
2,741
3,413
33_6
In this paper, we propose a cross-lingual pre-trained model (named as Xnlg) in order to transfer monolingual NLG supervision to other pre-trained languages by fine-tuning. Specifically, Xnlg shares the same sequence-to-sequence model across languages, and is pre-trained with both monolingual and cross-lingual objecti...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
6
3,413
3,987
33_7
The proposed model enables us to fine-tune the pre-trained model on monolingual NLG training data, and then evaluate it beyond a single language, including zero-shot cross-lingual generation. Besides, we explore several fine-tuning strategies to make a compromise between cross-lingual ability and task ability. In addi...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
7
3,987
4,500
33_8
Experimental results on the NLG tasks show that Xnlg achieves competitive performance compared with the machine-translation-based pipeline model in zero-shot cross-lingual settings. Related Work ::: Cross-Lingual NLG Several previous methods have been proposed for cross-lingual abstractive summarization. BIBREF7 xnh...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
8
4,500
5,046
33_9
However, the systems only conduct experiments that generate summaries with different languages from the input language, rather than transferring supervision signals across all language pairs. BIBREF10 kumar2019cross introduce a cross-lingual model for question generation, which uses training data annotated in multiple...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
9
5,046
5,507
33_10
Related Work ::: Monolingual Pre-Training Various training objectives are designed to pretrain text encoders used for general-purpose language representations, such as language modeling BIBREF11, BIBREF12, BIBREF13, BIBREF14, BIBREF15, auto-encoding BIBREF16, and machine translation BIBREF17. Apart from pre-training...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
10
5,507
6,011
33_11
Related Work ::: Cross-Lingual Pre-Training Cross-lingual pre-training aims at building universal cross-lingual encoders that can encode multilingual sentences to a shared embedding space. BIBREF20 artetxe2018massively use the sequence encoder of the multilingual translation model BIBREF3 to produce cross-lingual se...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
11
6,011
6,535
33_12
BIBREF4 xnli propose an alignment loss function to encourage parallel sentences to have similar representations. By pre-training BERT BIBREF13 on corpora of multiple languages, it shows a surprising ability to produce cross-lingual representations BIBREF21. More recently, BIBREF5 xlm extend mask language modeling pre-...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
12
6,535
7,128
33_13
Methods Xnlg is a pre-trained sequence-to-sequence model, which is based on Transformer BIBREF22. Both the encoder and the decoder are supposed to support multiple languages. Following BIBREF5, we use language tag embeddings to distinguish the source and target languages. Given a sentence and its corresponding langu...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
13
7,128
7,747
33_14
Methods ::: Pre-Training Tasks ::: Monolingual MLM The masked language modeling (MLM) BIBREF13 task, also known as the Cloze task BIBREF23, aims at predicting the randomly masked words according to their context. The objective pretrains the bidirectional encoder to obtain contextual representations. Following BIBREF...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
14
7,747
8,132
33_15
For each masked token, we substitute it with a special token M, a random token, or the unchanged token with a probability of 0.8, 0.1, and 0.1, respectively. Let $x$ denote a sentence from the monolingual training corpus, and $M_{x}$ the set of randomly masked positions. The monolingual MLM loss is defined as: MLM(x) ...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
15
8,132
8,535
33_16
Notice that language tags are fed into the model for all pre-training tasks. Methods ::: Pre-Training Tasks ::: Denoising Auto-Encoding (DAE) We use the denoising auto-encoding (DAE) objective BIBREF24 to pretrain the encoder-decoder attention mechanism. Given sentence $x$ from the monolingual corpus, we use three t...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
16
8,535
9,037
33_17
Third, we substitute tokens with the special padding token P with a probability of $0.1$. The pre-training objective is to recover the original sentence $x$ by conditioning on $\hat{x}$. The DAE loss is computed via: DAE(x) = -p(x|x) = -i = 1|x|p(xi | x, x<i) where $x_{<i}$ represents the tokens of previous time steps...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
17
9,037
9,380
33_18
Methods ::: Pre-Training Tasks ::: Cross-Lingual MLM (XMLM) Similar to monolingual MLM, the masked token prediction task can be extended to cross-lingual settings BIBREF5. To be specific, given a parallel corpus, we concatenate the pair of bilingual sentences $(x,y)$ to a whole sequence, and use it as the input of M...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
18
9,380
9,858
33_19
Apart from using monolingual context to predict the masked tokens, XMLM encourages the model to utilize the alignment of bilingual sentences, so that the model learns to map cross-lingual texts into a shared vector space. Similar to eq:mlm, the cross-lingual MLM loss is: XMLM(x,y) = -i Mxp( xi | xMx , yMy) -i Myp( yi...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
19
9,858
10,269
33_20
Methods ::: Pre-Training Tasks ::: Cross-Lingual Auto-Encoding (XAE) If only DAE is used as the pre-training task for the decoder, we found that the model ignores the target language tag while generating just the same language as the input, caused by the spurious correlation issue BIBREF25. In other words, the DAE l...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
20
10,269
10,780
33_21
To solve the above problem, we use machine translation as the cross-lingual auto-encoding (XAE) task, which decreases mutual information between the target sentences and the source language tag. XAE can be viewed as the multilingual-version DAE task in the sense that both of them recover the sentence by conditioning o...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
21
10,780
11,271
33_22
Methods ::: Pre-Training Protocol As shown in Figure FIGREF6(b), we propose a two-stage pre-training protocol for Xnlg. The first stage pretrains the encoding components, where the model learns to encode multilingual sentences to a shared embedding space. We consider using MLM and XMLM as the pre-training tasks.
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
22
11,271
11,587
33_23
The objective of the first stage is to minimize: 1= (x,y) p XMLM(x,y) + x m MLM(x) where ${_{\textnormal {p}}}$ indicates the parallel corpus, and ${_{\textnormal {m}}}$ is the monolingual corpus. Although the pre-trained encoder in the first stage enables the model to encode multilingual sentences.
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
23
11,587
11,889
33_24
However, it cannot directly be used in cross-lingual NLG because: 1) encoder-decoder attention is not pre-trained; 2) the decoding algorithm is different between masked language modeling and autoregressive decoding, resulting in the mismatch between pre-training and fine-tuning. Therefore, we conduct decoding pre-trai...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
24
11,889
12,339
33_25
The objective of the second stage is to minimize: 2 = (x,y) pXAE(x,y) + x mDAE(x) Methods ::: Fine-Tuning on Downstream NLG Tasks In the fine-tuning procedure, let us assume that we only have English training data for downstream NLG tasks. According to whether the target language is English, the directions of NLG ca...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
25
12,339
12,796
33_26
Methods ::: Fine-Tuning on Downstream NLG Tasks ::: Fine-Tuning for Any-to-Others NLG Ideally, the model can be fine-tuned towards a new task without losing its cross-lingual ability. However, we observe the catastrophic forgetting phenomenon of target language controllability, if we fine-tune all the model paramete...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
26
12,796
13,392
33_27
Methods ::: Fine-Tuning on Downstream NLG Tasks ::: Fine-Tuning for Any-to-English NLG For the Any-to-English NLG transfer, the decoder always generates English. So we can freeze the encoder parameters, and update the decoder parameters to retain the cross-lingual ability. As an alternative way, we can also fine-tun...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
27
13,392
13,825
33_28
Experiments We conduct experiments over two cross-lingual NLG downstream tasks, i.e., cross-lingual question generation, and cross-lingual abstractive summarization. We compare Xnlg with state-of-the-art cross-lingual pre-trained models, and machine-translation-based pipelines. Experiments ::: Training Details ::: ...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
28
13,825
14,336
33_29
In the first pre-training stage, we directly use the 15-language pre-trained XLM BIBREF5 to initialize the parameters of our encoder and decoder. In the second stage, we use Wikipedia as the monolingual data for the DAE objective, and MultiUN BIBREF27 as the parallel data for the XAE objective. The DAE loss is trained...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
29
14,336
14,817
33_30
Following BIBREF5, we use the tokenizer provided by BIBREF28 for Chinese, and Moses for other languages, respectively. Then the words in all languages are split with a shared subword vocabulary learned by BPE BIBREF29. We use Adam optimizer with a linear warm-up over the first 4,000 steps and linear decay for later st...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
30
14,817
15,258
33_31
It takes about 30 hours to run 23,000 steps for the pre-training procedure by using 4 Nvidia Telsa V100-16GB GPUs. Experiments ::: Training Details ::: Fine-Tuning For fine-tuning on downstream NLG tasks, we use Adam optimizer with a learning rate of $5\times 10^{-6}$. We set the batch size as 16 and 32 for question...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
31
15,258
15,734
33_32
When the target language is different from the language of training data, we fine-tune the Transformer layers of the encoder. We truncate the input sentences to the first 256 tokens. During decoding, we use beam search with beam size of 3, and limit the length of the target sequence to 80 tokens. Experiments ::: Ques...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
32
15,734
16,302
33_33
In the following experiments, we extend the QG task to the cross-lingual setting. By only using English QG training data, our goal is to generate questions in English or Chinese with the given passage-answer pair in English or Chinese. We use SQuAD 1.1 BIBREF30 as the English QG dataset. It is a popular English quest...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
33
16,302
16,721
33_34
Following BIBREF31, we regard the original development set as the test set, and sample 5000 examples from the training data of two datasets as the development sets. For Chinese QG, we follow the default data splits of WebQA BIBREF32. We regard the provided annotated evidence sentences as the input passages instead of ...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
34
16,721
17,243
33_35
During decoding Chinese, we utilize a subset of vocabulary, which is obtained from the passage sentences of the WebQA dataset. Experiments ::: Question Generation ::: English-English Question Generation We first conduct experiments on the supervised English-English QG setting. We compare our model to the following b...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
35
17,243
17,775
33_36
Xlm BIBREF5 The current state-of-the-art cross-lingual pre-training model. We initialize the Transformer-based sequence-to-sequence model with pre-trained XLM. We evaluate models with BLEU-4 (BL-4), ROUGE (RG) and METEOR (MTR) metrics. As shown in Table TABREF16, our model outperforms the baselines, which demonstrat...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
36
17,775
18,164
33_37
Experiments ::: Question Generation ::: Chinese-Chinese Question Generation We conduct experiments on the zero-shot Chinese-Chinese QG task to evaluate the cross-lingual transfer ability. In this task, models are trained with English QG data but evaluated with Chinese QG examples. We include the following models as ...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
37
18,164
18,723
33_38
We use the Transformer as the translator, which is also trained on the MultiUN dataset. Pipeline (Xlm) with Google Translator Same to Pipeline (Xlm) but using Google Translator to translate the texts. We evaluate models by both automatic evaluation metrics and human experts. The automatic metrics scores are computed...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
38
18,723
19,261
33_39
We randomly select 100 passage-answer pairs from the English QG test set, and use the models to generate questions. Then we present these examples to three experts to ask for the above scores. In Table TABREF17 and Table TABREF18, we present the results for the zero-shot Zh-Zh-QG. The results of monolingual supervised...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
39
19,261
19,775
33_40
In the human evaluation, our model also obtains significant improvements in terms of relatedness and correctness. Experiments ::: Question Generation ::: English-Chinese Question Generation In the zero-shot English-Chinese question generation experiments, we use Xlm and Pipeline (Xlm) as our baselines. Pipeline (Xlm...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
40
19,775
20,312
33_41
Table TABREF19 shows the human evaluation results, where our model surpasses all the baselines especially in terms of relatedness and correctness. Experiments ::: Question Generation ::: Chinese-English Question Generation We also conduct experiments for zero-shot Chinese-English question generation, and adopt the s...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
41
20,312
20,891
33_42
Experiments ::: Abstractive Summarization We conduct experiments on cross-lingual abstractive summarization (AS). AS is the task of converting the input sentences into summaries while preserving the key meanings. For evaluation, we use English/French/Chinese Gigaword to extract the first sentence and the headline of...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
42
20,891
21,377
33_43
Experiments ::: Abstractive Summarization ::: Zero-Shot Summarization In the zero-shot setting, we only use English data for training, and directly evaluate the model on other languages. In Table TABREF22 and Table TABREF23, we present the results for French/Chinese AS, which are evaluated by the ROUGE-1, ROUGE-2 an...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
43
21,377
21,877
33_44
Comparing with French, there is a larger gap between baselines and our model on zero-shot Chinese AS, which indicates that the error propagation issue is more serious on distant language pairs. Experiments ::: Ablation Studies ::: Effects of Pre-Training We conduct ablation studies for pre-training objectives, and t...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
44
21,877
22,489
33_45
Experiments ::: Ablation Studies ::: Effects of Fine-Tuning Strategies As shown in Table TABREF41, we use the En-En-QG and Zh-Zh-QG tasks to analyze the effects of using different fine-tuning strategies. It can be observed that fine-tuning encoder parameters, our model obtain an impressive performance for both Engli...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
45
22,489
23,047
33_46
We find that fine-tuning decoder hurts cross-lingual decoding, and the model learns to only decodes English words. For only fine-tuning decoder, the performance degrades by a large margin for both languages because of the underfitting issue, which indicates the necessity of fine-tuning encoder. Experiments ::: Ablati...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
46
23,047
23,600
33_47
Specifically, we first fine-tune Xnlg on the English AS data, and then fine-tune it on the French or Chinese AS data. We compare with the monolingual supervised model that Xnlg is only fine-tuned on the dataset of the target language. As shown in Figure FIGREF49, we can observe that the cross-lingual supervision impro...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
47
23,600
24,066
33_48
Experiments ::: Case Studies As shown in Figure FIGREF42, we present some examples generated by Xnlg and the baselines in four directions (En-En, En-Zh, Zh-En, and Zh-Zh). When decoding on an unseen language, Xlm tends to generate random output, because it is not designed for cross-lingual NLG. In terms of the pipel...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
48
24,066
24,551
33_49
For example, when the pipeline model performs Zh-Zh-QG, keywords are translated twice, increasing the risk of mistranslation. In the second example, “atomic bomb” is mistranslated to “nuclear bomb”, resulting in its low correctness. On the contrary, by directly transferring English supervision signals to the other gen...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
49
24,551
25,163
33_50
With the pre-trained model, we achieve zero-shot cross-lingual NLG on several languages by only fine-tuning once. Experimental results show that our model outperforms the machine-translation-based pipeline model on several cross-lingual NLG tasks. For future work, we would like to improve our pre-training method towar...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
50
25,163
25,723
33_51
Figure 2: Overview of the pre-training tasks and the pre-training protocol designed for XNLG. Table 3: Human evaluation results of zero-shot Chinese-Chinese question generation. Rel is short for relatedness, Flu for fluency, and Corr for correctness. “*” indicates the improvements are significant at p < 0.05. Table...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
51
25,723
26,252
33_52
Table 2: Evaluation results of zero-shot ChineseChinese question generation. Same shorthands apply as in Table 1. Table 4: Human evaluation results of zero-shot English-Chinese question generation. “*” indicates the improvements are significant at p < 0.05. Same shorthands apply as in Table 3. Table 5: Human evalua...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
52
26,252
26,731
33_53
Table 7: Evaluation results of zero-shot French abstractive summarization. Same shorthands apply as in Table 1. Table 6: Evaluation results of monolingual supervised abstractive summarization. Same shorthands apply as in Table 1. Table 8: Evaluation results of zero-shot Chinese abstractive summarization. Same short...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
53
26,731
27,240
33_54
Table 10: Effects of different fine-tuning strategies. Dec, Enc and ET represent fine-tuning the parameters of the decoder, encoder, and Transformer layers of encoder, respectively. Same shorthands apply as in Table 1. Figure 3: Examples of generated questions by XNLG and the baselines in four directions (En-En,En-Z...
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
54
27,240
27,705
33_55
Figure 4: ROUGE-2 scores for few-shot French/Chinese abstractive summarization with different training data sizes.
https://arxiv.org/abs/1909.10481
Cross-Lingual Natural Language Generation via Pre-Training
55
27,705
27,821
34_0
Hierarchical Neural Story Generation We explore story generation: creative systems that can build coherent and fluent passages of text about a topic. We collect a large dataset of 300K human-written stories paired with writing prompts from an online forum. Our dataset enables hierarchical story generation, where the m...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
0
0
606
34_1
Experiments show large improvements over strong baselines on both automated and human evaluations. Human judges prefer stories generated by our approach to those from a strong non-hierarchical model by a factor of two to one. Introduction Story-telling is on the frontier of current text generation technology: storie...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
1
606
1,181
34_2
We tackle the challenges of story-telling with a hierarchical model, which first generates a sentence called the prompt describing the topic for the story, and then conditions on this prompt when generating the story. Conditioning on the prompt or premise makes it easier to generate consistent stories because they pr...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
2
1,181
1,615
34_3
We find that standard sequence-to-sequence (seq2seq) models BIBREF1 applied to hierarchical story generation are prone to degenerating into language models that pay little attention to the writing prompt (a problem that has been noted in other domains, such as dialogue response generation BIBREF2 ). This failure is d...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
3
1,615
2,215
34_4
To improve the relevance of the generated story to its prompt, we introduce a fusion mechanism BIBREF3 where our model is trained on top of an pre-trained seq2seq model. To improve over the pre-trained model, the second model must focus on the link between the prompt and the story. For the first time, we show that fu...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
4
2,215
2,808
34_5
We improve efficiency using a convolutional architecture, allowing whole stories to be encoded in parallel. Existing convolutional architectures only encode a bounded amount of context BIBREF4 , so we introduce a novel gated self-attention mechanism that allows the model to condition on its previous outputs at differe...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
5
2,808
3,417
34_6
Experiments show that our fusion and self-attention mechanisms improve over existing techniques on both automated and human evaluation measures. Our new dataset and neural architectures allow for models which can creatively generate longer, more consistent and more fluent passages of text. Human judges prefer our hie...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
6
3,417
4,129
34_7
The prompts have a large diversity of topic, length, and detail. The stories must be at least 30 words, avoid general profanity and inappropriate content, and should be inspired by the prompt (but do not necessarily have to fulfill every requirement). Figure FIGREF1 shows an example. We scraped three years of prompts...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
7
4,129
4,769
34_8
We reserve 5% of the prompts for a validation set and 5% for a test set, and present additional statistics about the dataset in Table TABREF4 . For our experiments, we limit the length of the stories to 1000 words maximum and limit the vocabulary size for the prompts and the stories to words appearing more than 10 ti...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
8
4,769
5,246
34_9
As the dataset is scraped from an online forum, the number of rare words and misspellings is quite large, so modeling the full vocabulary is challenging and computationally intensive. Approach The challenges of WritingPrompts are primarily in modeling long-range dependencies and conditioning on an abstract, high-lev...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
9
5,246
5,888
34_10
We will evaluate strategies to address these challenges in the following sections. Hierarchical Story Generation High-level structure is integral to good stories, but language models generate on a strictly-word-by-word basis and so cannot explicitly make high-level plans. We introduce the ability to plan by decompos...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
10
5,888
6,493
34_11
Conditioning on the prompt makes it easier for the story to remain consistent and also have structure at a level beyond single phrases. Efficient Learning with Convolutional Sequence-to-Sequence Model The length of stories in our dataset is a challenge for RNNs, which process tokens sequentially. To transform prompt...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
11
6,493
7,095
34_12
In the Conv seq2seq model, the encoder and decoder are connected with attention modules BIBREF7 that perform a weighted sum of encoder outputs, using attention at each layer of the decoder. Modeling Unbounded Context with Gated Multi-Scale Self-attention CNNs can only model a bounded context window, preventing the m...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
12
7,095
7,660
34_13
The self-attention mechanism improves the model's ability to extract long-range context with limited computational impact due to parallelism. Gated Attention: Similar to BIBREF9 , we use multi-head attention to allow each head to attend to information at different positions. However, the queries, keys and values are ...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
13
7,660
8,206
34_14
Multi-Scale Attention: Further, we propose to have each head operating at a different time scale, depicted in Figure FIGREF7 . Thus the input to each head is downsampled a different amount—the first head sees the full input, the second every other input timestep, the third every third input timestep, etc. The differe...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
14
8,206
8,690
34_15
The output of a single attention head is given by DISPLAYFORM0 where INLINEFORM0 contains the hidden states up to time INLINEFORM1 at layer INLINEFORM2 , and INLINEFORM3 are gated downsampling networks as shown in Figure FIGREF7 . Unlike BIBREF9 , we allow the model to optionally attend to a 0 vector at each times...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
15
8,690
9,233
34_16
Additionally, we do not allow the self-attention mechanism to attend to the current timestep, only the past. Improving Relevance to Input Prompt with Model Fusion Unlike tasks such as translation, where the semantics of the target are fully specified by the source, the generation of stories from prompts is far more ...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
16
9,233
9,866
34_17
We train a seq2seq model that has access to the hidden states of a pretrained seq2seq model. Doing so can be seen as a type of boosting or residual learning that allows the second model to focus on what the first model failed to learn—such as conditioning on the prompt. To our knowledge, this paper is the first to sho...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
17
9,866
10,333
34_18
The cold fusion mechanism of BIBREF3 pretrains a language model and subsequently trains a seq2seq model with a gating mechanism that learns to leverage the final hidden layer of the language model during seq2seq training. We modify this approach by combining two seq2seq models as follows (see Figure FIGREF13 ): DISPL...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
18
10,333
10,907
34_19
The gated hidden layers are combined by concatenation and followed by more fully connected layers with GLU activations (see Appendix). We use layer normalization BIBREF10 after each fully connected layer. Story Generation Sequence-to-sequence neural networks BIBREF1 have achieved state of the art performance on a va...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
19
10,907
11,465
34_20
Previous work on story generation has explored seq2seq RNN architectures BIBREF14 , but has focused largely on using various content to inspire the stories. For instance, BIBREF15 uses photos to inspire short paragraphs trained on romance novels, and BIBREF16 chain a series of independent descriptions together into a...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
20
11,465
12,122
34_21
They find this technique can generate text of the desired genre, but the movie plots are not interpretable (as the model outputs events, not raw text). However, we are not aware of previous work that has used hierarchical generation from a textual premise to improve the coherence and structure of stories. Hierarchica...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
21
12,122
12,745
34_22
BIBREF20 generate a discrete latent variable based on the context, then generates text conditioned upon it. Fusion Models Previous work has investigated the integration of language models with seq2seq models. The two models can be leveraged together without architectural modifications: BIBREF21 use language models t...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
22
12,745
13,295
34_23
BIBREF23 combined a trained language model with a trained seq2seq model to learn a gating function that joins them. BIBREF3 propose training the seq2seq model given the fixed language model then learning a gate to filter the information from the language model. Baselines We evaluate a number of baselines: (1) Langu...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
23
13,295
13,824
34_24
(2) seq2seq: using LSTMs and convolutional seq2seq architectures, and Conv seq2seq with decoder self-attention. (3) Ensemble: an ensemble of two Conv seq2seq with self-attention models. (4) KNN: we also compare with a KNN model to find the closest prompt in the training set for each prompt in the test set. A TF-IDF...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
24
13,824
14,247
34_25
The retrieved story from the training set is limited to 150 words to match the length of generated stories. Fusion Training To train the fusion model, we first pretrain a Conv seq2seq with self-attention model on the WritingPrompts dataset. This pretrained model is fixed and provided to the second Conv seq2seq with ...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
25
14,247
14,768
34_26
Similar to BIBREF6 , we train using the Nesterov accelerated gradient method BIBREF26 using gradient clipping BIBREF27 . We perform hyperparameter optimization on each of our models by cross-validating with random search on a validation set. We provide model architectures in the appendix. Generation We generate stor...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
26
14,768
15,427
34_27
We find this sampling strategy substantially more effective than beam search, which tends to produce common phrases and repetitive text from the training set BIBREF28 , BIBREF29 . Sentences produced by beam search tend to be short and generic. Completely random sampling can introduce very unlikely words, which can dam...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
27
15,427
16,018
34_28
To ease human evaluation, we generate stories of 150 words and do not generate unknown word tokens. For prompt generation, we use a self-attentive GCNN language model trained with the same prompt-side vocabulary as the sequence-to-sequence story generation models. The language model to generate prompts has a validati...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
28
16,018
16,645
34_29
Many commonly used metrics, such as BLEU for machine translation or ROUGE for summarization, compute an n-gram overlap between the generated text and the human text—however, in our open-ended generation setting, these are not useful. We do not aim to generate a specific story; we want to generate viable and novel stor...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
29
16,645
17,165
34_30
Perplexity is commonly used to evaluate the quality of language models, and it reflects how fluently the model can produce the correct next word given the preceding words. We use prompt ranking to assess how strongly a model's output depends on its input. Stories are decoded under 10 different prompts—9 randomly sampl...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
30
17,165
17,775
34_31
For human evaluation, we use Amazon Mechanical Turk to conduct a triple pairing task. We use each model to generate stories based on held-out prompts from the test set. Then, groups of three stories are presented to the human judges. The stories and their corresponding prompts are shuffled, and human evaluators are a...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
31
17,775
18,357
34_32
We use Amazon Mechanical Turk to compare the stories from hierarchical generation from a prompt with generation without a prompt. 400 pairs of stories were evaluated by 5 judges each in a blind test. Results We analyze the effect of our modeling improvements on the WritingPrompts dataset. Generation Quality Our pr...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
32
18,357
18,897
34_33
In contrast, the baseline Conv seq2seq model copies 10.2 words on average and the KNN baseline copies all 150 words from a story in the training set. Figure FIGREF27 shows the values of the fusion gates for an example story, averaged at each timestep. The pretrained seq2seq model acts similarly to a language model pr...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
33
18,897
19,433
34_34
For example, can't is tokenized to ca n't, and the model occasionally produces the first token but misses the second. A similar error is after one line of dialogue, the model may move to another line of dialogue without generating a newline token. A further obstacle is repetition. The model focuses frequently on what ...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
34
19,433
19,968
34_35
Language models often struggle to model rare words accurately, as the probability distribution over the next word is dominated by more common words. This tends to produce similar prompts, particularly at the start — we see many prompts that start with the man. In contrast, many of the human prompts are very unique (e....
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
35
19,968
20,459
34_36
Use of Attention We analyze the encoder-decoder attention in the fusion model and find that unlike attention maps in machine translation, where each decoder timestep tends to attend to a different word on the encoder-side, the attention map for each decoder timestep looks similar and focuses mainly on salient words ...
https://arxiv.org/abs/1805.04833
Hierarchical Neural Story Generation
36
20,459
21,125